Skip to content

progress.md

Latest commit

 

History

History
75 lines (69 loc) · 9.8 KB

progress.md

File metadata and controls

75 lines (69 loc) · 9.8 KB
Raw

Progress

2026-06-06 Settings Preflight + 100-PDF Live Regression

  • Started current task at 2026-06-06 17:43:02 CST.
  • User provided a test-only OpenAI-compatible API key and endpoint https://icoe.pp.ua/; key will be used only as a transient environment variable and not written to repo files.
  • Read frontend-skill and planning-with-files instructions for this UI + long regression task.
  • Confirmed Settings preflight UI is currently bulky because every preflight check is rendered as a full .layer-list card.
  • Confirmed /Users/maziheng/Downloads/0.3.1 working/ReadingPractice/PDF contains 262 PDFs.
  • Confirmed ReadingPractice has no legacy JS oracle directory, so a new no-legacy 100-PDF Live pipeline diagnostic script is needed.

2026-06-06 Windows Package and Compatibility

  • Started Windows package and compatibility goal from existing active /goal.
  • Used planning-with-files workflow because the task requires a persistent, trackable task record across many implementation steps.
  • Session catchup skipped because Codex native session parsing is not implemented by the helper script.
  • Read Files/Windows包体与兼容规划.md and split the work into W0-W10.
  • Detected dirty worktree before edits; no existing changes were reverted.
  • Spawned three concurrent Subagents:
    • release/audit worker for package.json, package audit, macOS path compatibility, offline WebView2 config.
    • runtime worker for Rust Python resolver, PDF renderer adapter, environment preflight.
    • dev/e2e worker for Windows-compatible UI flow, preview, and PDF regression scripts.
  • Created Files/Windows包体与兼容任务追踪.md.
  • Updated Files/Windows包体与兼容规划.md with an execution tracking index.
  • Added the Windows compatibility active goal to the top of task_plan.md, findings.md, and progress.md while preserving prior task history.
  • W6 dev/e2e Subagent completed: ui-flow-e2e now uses fileURLToPath, Windows npm.cmd, and Windows Chrome/Edge paths; preview-e2e Python resolver supports EPIC8_PYTHON/Windows py -3; pdf-regression-sample no longer has a local macOS corpus default and supports Windows .exe.
  • W6 validation reported passed: node --check for the three scripts, node scripts/pdf-regression-sample.mjs --help, expected exit 2 without explicit corpus dirs, and npm run check.
  • Sent W6 integration finding to release/audit Subagent: package.json must update test:pdf-regression because the script now requires explicit --pdf-dir and --legacy-dir.
  • W1/W2 release/audit Subagent completed: split macOS/Windows release scripts, platformized package audit, added Windows artifact metadata/sha256/WebView2/lockfile reporting, fixed repack-macos-dmg.mjs path handling, and added src-tauri/tauri.windows.offline.conf.json.
  • W1/W2 validation reported passed: npm run check, node scripts/package-audit.mjs --help, npm run audit:package, expected Windows missing-artifact failure, fake Windows artifact success branch, expected npm run test:pdf-regression exit 2 without corpus dirs, and git diff --check.
  • W8 remains open because Authenticode/PowerShell signature audit was not added in the release/audit worker pass.
  • W3-W5 runtime Subagent completed and parent fixed one Rust borrow-check issue in merge_rendered_page_images.
  • Runtime validation passed: cargo check, cargo test environment_preflight_reports_required_dependency_names, and cargo test pdf_render_adapter.
  • Added scripts/audit-windows-signatures.ps1, scripts/windows-install-instructions.txt, and .github/workflows/windows-smoke.yml to close W7-W9.
  • Final local verification passed: npm run check, cargo fmt --check, cargo check, cargo test --manifest-path src-tauri/Cargo.toml (115 passed, 2 ignored), git diff --check, node --check for the changed JS scripts, node scripts/package-audit.mjs --help, and npm run audit:package.
  • Expected local limitations recorded: npm run audit:package:windows fails because this macOS workspace has no NSIS/MSI artifact; pwsh is unavailable locally for the Windows signature script; npm run test:pdf-regression now intentionally requires explicit --pdf-dir and --legacy-dir.

2026-06-04

  • Started validation from existing dirty worktree.
  • Confirmed key files are modified and scripts/pdf-regression-sample.mjs exists.
  • npm run check passed.
  • cargo test --manifest-path src-tauri/Cargo.toml passed: 94 passed, 2 ignored.
  • jq is not installed locally; switched JSON inspection to Node one-liners.
  • Fixed dev fallback routing so clear-text PDF lands directly on the editable draft page.
  • npm run e2e:ui-flow passed with real fixture PDFs; scanned/manual transcription flow now creates a new draft revision instead of being blocked by overwrite protection.
  • Added Rust regression coverage for manual transcription replacing the source document, archiving the previous draft, and regenerating a new editable draft.
  • Tightened classifier rules after inspecting the 30-PDF regression report: explicit A-D evidence is required for single choice, explicit Choose TWO/THREE letters is required for multi choice, according to no longer triggers classification, Questions X and Y now preserves both question ids, and match each statement/opinion/person is ordinary matching rather than paragraph information matching.
  • Updated PDF regression normalization so legacy JS table-rendered paragraph matching and old completion display variants compare by semantic kind instead of raw legacy field names.
  • Changed auto-pipeline routing so generated drafts open the editable draft page by default; only source-document review keeps the user on document confirmation.
  • Fixed remaining random-regression structure issues: sentence-ending matching now wins over single-choice detection, overlapping ranges are normalized, flow/diagram completion uses inline completion layout, and explicit completion groups can extend to numbered blanks present in the same source span.
  • Final pre-build checks passed: npm run check, cargo test --manifest-path src-tauri/Cargo.toml (98 passed, 2 ignored), npm run e2e:ui-flow, and npm run test:pdf-regression with 30/30 structure pass on seed 1780508509492.
  • Random regression answer comparison still reports missing answers when answer pages have no extractable text; this is tracked as parser/scan limitation and is routed to user confirmation rather than production overwrite.
  • Built fresh DMG: /Users/maziheng/Downloads/Desktop/copy/PDF2Test/src-tauri/target/release/bundle/dmg/IELTS Author Studio_0.1.0_aarch64.dmg, size 5.7 MB, SHA-256 278d89b830ff5021f25bfd6918f9aa2358a77ce6456ec49606f01a28fa615a7d.
  • Re-ran the real OpenAI-compatible API chain with the new test profile on the Margaret Preston PDF; the CLI completed successfully and wrote tmp/live-auto-pipeline-margaret-new-key.json.
  • Real vision answer extraction succeeded with 13 answers and populated q8-q13 as symbols, titles, stencilling, books, travel, and 400.
  • Real cloud whole-paper comparison was attempted. Direct PDF upload was rejected by the provider as unsupported, the image fallback returned JSON, and the local draft remained authoritative because the cloud comparison disagreed with the local 8-13 layout/kind.
  • Verified the resulting local draft status is Authoring, nextRoute=groups, Questions 8-13 is one sentence_completion group with layoutHint=inline_completion, and q8-q13 remain in the same group.
  • npm run check passed after updating the routing tests.
  • cargo test --manifest-path src-tauri/Cargo.toml passed: 98 passed, 2 ignored.
  • npm run e2e:ui-flow failed because the script still expects the old source-review route during the OCR flow; product behavior now routes generated drafts to groups.
  • npm run test:pdf-regression completed on seed 1780555370790: 27/30 structure pass, 3 structure failures, 30/30 answer failures due to parser limitations on answer pages.

2026-06-06

  • Read the pasted export failure: publish gate was blocked by NeedsReview, source-review parser warnings for page 5/6 no extractable text, low-confidence groups/questions, and empty answers.
  • Added import UI stage tracking. During runAutoPipeline, the user now sees a cloud model wait message rather than only "正在生成".
  • Added pipeline report fields and persisted audit summaries for vision answer extraction: attempted/applied state, answer count, filled question ids, and missing question ids.
  • Added cloud comparison report fields and persisted audit summaries: issues, observations, local outline summary, and cloud outline summary. Local authoring remains authoritative.
  • Added editor UI blocks for "视觉答案补全" and "云端整卷对照", including local-vs-cloud outline summaries and missing-answer warnings.
  • Added "确认当前题组" action in the group editor to mark all questions in the active group verified after the user checks them.
  • Reworked export-page error handling to parse *_validation_failed:{...} JSON payloads and show actionable guidance for empty answers, source review, verification, and cloud comparison.
  • Extended the Node LLM sidecar command whitelist with extract_pdf_image_answers and generate_pdf_reading_outline for parity with the Rust gateway.
  • Updated Rust cloud/vision regression coverage: cloud comparison summary persists after minimization, cloud mismatch does not overwrite local answers, and vision answer extraction exposes filled/missing question ids.
  • Updated PDF image extraction tests to match full-page vision rendering behavior rather than the older single sips preview assumption.
  • Verification passed: npm run check; node --check sidecars/llm-gateway/gateway.mjs; cargo check --manifest-path src-tauri/Cargo.toml; cargo test --manifest-path src-tauri/Cargo.toml (114 passed, 2 ignored).
  • Started Vite at http://127.0.0.1:1420/ after approval and checked the import wizard with the in-app browser.